Overview

• One-SHOT retrieval proposes adaptive chunk selection under a token budget with additional filtering to pack more relevant material into context windows.
• The same paper tests an iterative, agentic RAG setup where a reasoning model issues and refines queries over multiple turns to counter query drift and retrieval laziness in government and regulatory corpora.
• VaccineRAG introduces a Chain-of-Thought–based dataset that prompts models to analyze each retrieved sample before answering, improving discrimination of harmful or irrelevant items.
• The VaccineRAG framework adds Partial-GRPO to help models learn long, structured CoT outputs, with authors reporting effectiveness in evaluations and pledging to release code and data.
• A CFA-focused benchmark evaluates 1,560 official mock questions and finds reasoning-oriented models strongest in zero-shot, while a hierarchical, structured RAG pipeline boosts accuracy and highlights knowledge gaps as the main failure mode.